The ACM Recommender Systems conference opened on 15 September 2016 in Boston (MA) with the RecSys TV workshop, one day dedicated to recommender systems in the broadcasting sphere.
Presentations by Verizon, Comcast, Netflix, the University of Lisbon were the highlights of the day. Diana Hu (Verizon) showed how time-series can be used to predict users’ preferences. An interesting decomposition in genres showed very recurrent patterns throughout the week. This, added to other recurring patterns on longer time scales, showed that we are creatures of habits. The best way to predict right is, Diana Hu concluded, to cope with those habits.
Then came the presentation by Netflix, given by Sr. Research Engineer Hossein Taghavi. The presentation gave excellent insights into the Netflix machinery, especially how they manage to keep the customer watching. Hossein explained how they conceived the Continue Watching (CW) function, one of many useful algorithms at Netflix. CW algorithms predict what you are likely to resume watching. CW prediction algorithms customize the position of the titles within a single row but also the position of the CW row itself. This functionality, like all other recommendation systems on Netflix, aims at maximizing one (and only one) metrics: customer retention. Like Hossein explained, there is no other metrics that counts at Netflix. Everything flows from there. Recommender systems aims at increasing the time spent using Netflix, hence reinforcing habits and creating expectations (and hopefully satisfaction). Those habits and expectations are then transformed into renewed subscriptions or, in other words, customer retention.
Another presentation worth mentioning was given by Miguel Costa (University of Lisbon). Miguel used databases of 682m views of Live TV, Catch UP and VOD to research how behaviors differ between those 3 types of content. One very interesting conclusion is that the percentage of content consumed varies according to the type of content. The percentage of content watched on Live TV is small (27%) when compared with Catch-Up (50%) and VOD (60%). Another conclusion is that kids watch content more “completely” than adults. This is very intuitive. Look at your kids and you’ll note that they indeed won’t stop watching a cartoon unless you force them to. Eventually Miguel Costa concluded rightfully that recommendations should take the context (Live, VOD, Catch-UP) into account (whereas you may object that Live TV doesn’t need recommendations, which is entire discussion by itself).
Posted in big data, Innovation, Marketing.